A statistical image model for motion estimation

Model-based object-oriented motion estimation from image sequences is addressed. A generic label field segments the scene into several continuously moving 2-D objects. An image model assuming segmentwise stationarity of the displaced frame difference (dfd) and of the estimated fields is proposed. The dfd is shown to obey a white generalized Gaussian distribution better than the commonly assumed overall white Gaussian distribution. A coupled weak smoothness constraint bounds the segments of the label field to smooth shape and the vector field to smoothness within each of those segments. A MAP (maximum a posteriori) estimator with respect to the image model is derived. Its performance is demonstrated by experimental results.<<ETX>>